TLDR AI 2023-10-31

Researchers wrote that gpt-3.5-turbo (the model behind ChatGPT) is a 20B parameter model in a paper by a Microsoft research group that studies diffusion models for code. The original model was said to be 175B parameters, so there could have been retraining or distillation.

The group behind Stable Diffusion has launched the Open Empathic project to foster emotion-detecting capabilities in open-source AI systems for more authentic AI-human interactions.

President Joe Biden signed an executive order outlining eight goals to regulate generative AI, focusing on safety, privacy, equity, consumer protection, innovation, and responsible government use, with agencies like the NIST tasked to develop AI standards and large AI model developers required to share safety test results.

A post from MosaicML on using AMD GPUs for training large language models.

The HuggingFace team reproduces a 2019 repository from OpenAI that introduced some of the first RLHF ideas.

Embeddings are machine-readable lists of numbers that represent some information about a piece of data. They are increasingly obtained by running the data through a neural network and looking at the activations of the model. This post pushes that idea forward with CLIP on a computer vision task and explores how to use dimensionality reduction to reason about the learned embeddings.

Transcribe 300 minutes of audio in less than 10 minutes with OpenAI’s Whisper Large v2.

This paper introduces Image Clustering Conditioned on Text Criteria (IC⁠TC), a method that leverages vision-language models to sort images based on user-given text descriptions.

This project aims to create a simple and easy to deploy solution to add observability to your Azure OpenAI instance. The approach adds an API Management instance as a proxy for your existing Azure OpenAI service, and funnels logs/requests/responses to an Application Insights instance. It also saves a prebuilt query to a workbook for easy access to logs.

Qualcomm's upcoming Snapdragon 8 Gen 3 chip, set to debut in Samsung’s Galaxy S24, emphasizes advanced AI camera tools and enhanced processing speeds.

Apple is lagging in AI advancements, especially when compared to competitors like Google and Microsoft, and there are doubts among Apple insiders that its AI/ML team can deliver.

AI pioneers Yoshua Bengio and Geoffrey Hinton, along with 22 other AI experts, have proposed a policy framework emphasizing AI safety and have recommended that companies and governments allocate a third of their AI R&D budget to safety efforts.

Easily integrate Large Language Models into your Python code.

Realtime UI personalization, powered by AI.

The infrastructure layer in AI is rapidly changing due to accelerated hardware development, fierce optimization competition, fast-paced research, long feedback cycles, ambiguity in other layers, and the commoditization of infrastructure components.

Labelbox is a platform for building data-centric AI applications. As the Foundation Models Lead, you will lead research and development for applying foundation models to real world problems for customers like Walmart and Adobe. Learn more.